An Updated Treatment of Aerosol Optical Properties in E3SM using a Neural Network Mie Scattering Emulator

Atmospheric aerosols strongly impact the Earth’s radiative budget and thus climate. Their primary climate impact is through interactions with clouds, but their direct radiative effects are also substantial, particularly for strongly absorbing aerosols like black carbon. In the Department of Energy’s Energy Exascale Earth System Model (E3SM) aerosol optical properties are parameterized using an approach similar to a look-up table that contains the bulk optical properties of aerosol populations predicted using Mie theory. Because the parameterization has a high-dimensional input-space the resolution of the table, and thus the accuracy, is limited. Furthermore, the current parameterization assumes that different aerosol species are internally mixed within the four size modes that are used by the Modal Aerosol Module in E3SM, which can lead to errors in parameterized optical properties. This work develops a neural network emulator that is trained on a new Python-based Mie code to predict the bulk optical properties of aerosol populations. It is capable of representing particles both as homogeneous spheres and concentric spheres of different materials. The new emulator has negligible error (0.08% mean absolute percentage error) in its predicted optical properties. We compare several climate simulations using the new parameterization and show how simulated aerosol radiative effects are altered by 1) using the new parameterization to achieve more accurate bulk aerosol optical properties and 2) representing mixed black carbon and sulfate aerosol using a core-shell particle model instead of assuming homogenous mixing.